Advancing Burn Patient Care: AI-Driven Prognosis Models

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Sep 20 2023

Burn injuries, spanning from minor inconveniences to life-threatening situations, necessitate tailored treatment based on the severity and location. In a recent publication in the journal Diagnostics, researchers introduced versatile machine-learning models designed to forecast burn patient outcomes, including the need for graft surgery and prolonged hospitalization.

Background

Burn injuries encompass damage to the skin or underlying tissues resulting from exposure to fire, heat, electricity, chemicals, or radiation. They range from mild superficial burns to life-threatening deep burns. Burns are classified into different degrees based on severity and depth:

• First-degree burns: superficial, affecting only the epidermis, causing pain, redness, and minor swelling.
• Second-degree burns, affecting both the epidermis and dermis, are characterized by severe pain, blistering, swelling, and redness.
• Third-degree burns: extending through all skin layers, often appearing white, charred, or blackened, with nerve damage necessitating surgical intervention.

Burn severity assessment includes using the "Rule of Nines," dividing the body into regions, each signifying the total body surface area (TBSA) percentage. It aids in gauging the burn's extent and deciding on the necessity of specialized care. Immediate first aid involves heat removal and cool water application. Severe burns demand medical attention, including wound care, topical medications, and potential skin grafts.

For precise burn assessment, the Baux score factors in age and TBSA affected. A higher score indicates increased mortality risk, aiding treatment decisions. Yet, its accuracy varies, prompting exploration of alternative scores such as body mass index (BMI) and the Anesthesiologist Physical Status Score.

Severe burns demand real-time monitoring, driving the integration of AI to enhance clinical care quality. The present study aims to develop a high-risk burn prediction model using ML approaches to meet modern precision medicine requirements and enhance burn patient outcomes.

Machine learning models for burn patients

The study was conducted at Chi Mei Medical Center, and it encompassed all hospitalized individuals with burn injuries, excluding those less than six years old. A total of 348 cases from January 1, 2010, to December 31, 2019, were analyzed. Ethical approval was granted by the Institutional Review Board of Chi Mei Medical Center, with patient consent waived due to the retrospective nature.

Three outcome variables were selected for prediction models: graft surgery, prolonged hospitalization (more than 14 days), and overall adverse effects (sepsis, pneumonia, respirator use, mortality, chronic kidney disease, and prolonged hospitalization). Twelve feature variables were used for model construction based on literature and input from clinical expertise. These included BMI, gender, age, smoking history, burn area, escharotomy, burn site (perineum and extremities), and lab results (white blood cell, creatinine, hemoglobin, and glutamate pyruvate transaminase).

The models incorporated all variables and were divided into a training dataset and a testing dataset. The synthetic minority oversampling technique (SMOTE) addressed the data imbalance. Four machine-learning algorithms (logistic regression, random forest, LightGBM, and XGBoost) were employed. For optimizing model performance, researchers employed the grid search technique with 5-fold cross-validation. Evaluation metrics included accuracy, specificity, sensitivity, and the area under the curve (AUC).

Results and analysis

Demographically, burn patients had an average age of 45.8 years, with 50.4 percent categorized as having the least severe burn area rank and 18.8 percent falling into the most severe category. The average Baux score stood at 69.2.

Machine-learning models showed varying performance. In graft surgery prediction, the random forest model had the highest AUC (0.757), followed closely by logistic regression, LightGBM, and XGBoost models. XGBoost excelled with an AUC of 0.815 for prolonged hospital stays, outperforming random forest, LightGBM, and logistic regression. In the case of overall adverse effects, LightGBM led with an AUC of 0.845, surpassing logistic regression, random forest, and XGBoost.

Feature importance was assessed through SHAP (SHapley Additive exPlanations) analysis, which highlighted the key contributors to each outcome. The AI models significantly outperformed the Baux score in predicting prolonged hospital stays and overall adverse effects. Furthermore, an AI risk prediction system was developed and received positive feedback from pilot burn care staff, ultimately enhancing the quality and efficiency of burn patient care.

Conclusion

In summary, the authors aimed to create a versatile machine-learning model for burn patient prognosis. They seamlessly integrated the model as an online predictive application into the hospital information system without the need for complex operations. The researchers believe that using machine-learning algorithms to predict outcomes in burn patients can help physicians promptly determine disease severity after hospital admission, which enables them to choose the most ideal and personalized treatment options, thus significantly improving prognosis. Physicians welcomed this integration, highlighting its potential to boost patient outcomes. Future research should explore additional variables and feature selection, harnessing evolving diagnostic tools for more data and better system performance.